The three things between your business and AI that works

Somewhere in the last two years, most small and mid-sized companies tried an AI tool. Usually a chatbot. The demo was impressive, the setup took an afternoon, and for a few weeks it answered questions on the website. Then a customer asked it something real, it gave a confident wrong answer, and someone quietly switched it off. The owner drew a reasonable conclusion: AI doesn’t work for a business our size.

The conclusion is wrong, and the research on why is more useful than it sounds. The chatbot was the visible tenth of a system. Three other parts were never built: a connection to the business’s real data, a handoff to a person, and someone whose job it was to watch it. All three are ordinary engineering. Here’s how the businesses that get value from AI put them in place, and how we do it.

RAND reached a matching conclusion from the engineering side. Its interviews with 65 experienced AI engineers found the leading cause of failure was that nobody had agreed what problem the project was meant to solve. Missing data came second, then chasing the technology for its own sake, then missing infrastructure. Four of the five root causes are decisions rather than code, which is what makes them fixable.

Ask an owner what the chatbot knew, and the honest answer is usually the FAQ page. It couldn’t check whether Tuesday at three was free, whether an invoice had been paid, or whether the part was in stock. That information lives in a booking tool, an accounting system, a spreadsheet, and someone’s head. The bot could only recite, and customers worked that out in one conversation.

The second gap is the one customers notice first. If the bot can’t answer a question, most people will forgive it once. If it can’t answer and there’s no way to reach a person, they leave. What they want is simple: let the assistant handle the easy things, and open a door to a real person the moment it can’t.

Klarna learned this in public. In February 2024 it said its AI assistant was doing the work of 700 agents. By May 2025 its CEO was telling Bloomberg that cost had been too dominant a factor, that the result was lower quality, and that the company was hiring people back. Klarna kept the AI running and put humans back on the cases it couldn’t handle. That split is the one to copy at any size, and Gartner’s numbers show where the line falls: back in 2023, chatbots resolved only 17% of billing disputes.

So we build the handoff first. The assistant knows the edge of what it can do. When a question falls outside that edge, or the customer asks for a person, the whole transcript lands with a named human in a channel they already read: the inbox, a Slack channel, a text message. The customer never repeats themselves, and the person picking up already knows what was tried.

The third gap is the quietest. The chatbot was bought by whoever saw the ad, set up in an afternoon, and from then on belonged to no one. Nobody read the transcripts. Nobody updated it when prices changed in March. Nobody noticed it had spent a month telling people you’re open on Sundays. A tool that talks to customers needs an owner the way a till needs someone to cash it up.

Our answer is a one-page agreement before anything is built: the one job the assistant does, what a good answer looks like, and who on your side owns it. Then a monthly look at the transcripts together, so what it got wrong last month is fixed this month. Prices change, policies change, the assistant changes with them. That loop is what separates a tool still running in a year from one switched off in week six.

Put the three fixes together and a second attempt looks less like a chatbot and more like a small piece of infrastructure. It starts with a short review of your operations to find the one job worth doing first, which is often behind the counter rather than in front of it. From there the build has four parts:

• One narrow job it can finish end to end. Rescheduling appointments, answering order-status questions, sorting the shared inbox. Not “customer service.”
• Read access to the systems that hold the answers. Booking tool, invoices, stock list, so it answers from your data instead of a text box you filled in once.
• A handoff that carries the full conversation to a named person, in a channel they already check.
• An owner and a monthly review, so it improves instead of drifting.

The model is the easy part; you can swap it in an afternoon. The connections, the escape hatch, and the person watching it are the system, and they’re what you own when we’re done. Build those and the chatbot mostly takes care of itself.